Traditional image interpolation methods always are introduced by piecewise-cubic kernel, which have been derived in ID with one parameter and applied to 2D images in a separable way. However, images typically are statistically nonseparable, and it motivates the construction of nonseparable cubic kernel function. This paper derives a new nonseparable cubic kernel based on the Coons-type surface method, then the efficiency of our method is compared with that of the traditional algorithm of cubic kernel for image resizing in Matlab 7. The experimental results show that our proposed algorithm excels the bicubic interpolation method in visual effect and complexity.